ArXiV ML/AI/CV papers summary

Theme 1: Mechanistic Interpretability and Structural Dynamics

The field is moving beyond “black-box” scaling toward a rigorous understanding of internal model geometry and optimization. We are shifting from aggregate performance metrics to examining the non-linear manifolds and symmetry-dictated representations that govern model behavior.

• Internal Geometry: Research is uncovering that concepts in LLMs are often non-linear manifolds rather than simple linear directions, as seen in Beyond Linear Concepts: Discovering and Aligning Non-Linear Concept Manifolds in Large Language Models. Similarly, Symmetry Discovery in Quantum Learning: Observable-Level and Task-Level Inference from Finite Measurements highlights how group symmetries dictate the geometry of learned representations.
• Causal Auditing: Skepticism toward “self-repair” is rising; Every Ablation Is a Dose: Counterweights and the Semblance of Self-Repair suggests that apparent self-repair is often just the activation of pre-existing “counterweights.” Furthermore, Are We Recovering Mechanisms? Objective-Level Recovery Gaps in Mechanistic Interpretability warns of a “recovery gap” where discovery algorithms fail to align with the objectives used to evaluate them.
• Persistent Organization: Persistent Depth Ordering amid Shifting Block-Bypass Responses in Language Model Pretraining provides a longitudinal view, proving that layer sensitivity is dynamic and that single-checkpoint analyses can be misleading.

Theme 2: Agentic Reasoning, Reliability, and Verification

As agents transition from research prototypes to enterprise tools, the focus has shifted from simple success rates to evidence-based verification and “assurance by construction.”

• Verification Frameworks: We are moving toward rigorous, evidence-bound execution. VeriHarness: Scaling Agentic Verification for Long-Horizon Tasks and Praxa, an Evidence-Bound Harness for Governed AI Agent Execution emphasize that we must verify the evidence behind an agent’s claim. This is supported by Actions with Receipts: Jointly Binding Claims, Evidence, and Execution for Replayable Tool-Agent Auditing and Verify Claims, Not Scores: Evidence-Based Verification of Modular Agents.
• Debugging and Attribution: DeFA: Dependency-Guided Failure Attribution for LLM Agents and When Harnesses Lose the Signal: Causal Evaluation of Recovery in LLM Agents provide frameworks to trace failures to their source, while My FAULT: Self-Diagnosis as Credit Assignment in Self-Evolving Agentic Reinforcement Learning turns error detection into a learning signal.
• Multi-Agent Coordination: Global Coherence: When Every Agent Is Right and the Team Is Still Wrong - A Local-to-Global Semantic Foundation for Multi-Agent Collaboration and Worse Together: How Performance Breaks Down in Multi-User Multi-Agent Teams highlight the “Global Coherence Problem,” where local intelligence fails to overcome missing global state.

Theme 3: Physics-Informed and Embodied Intelligence

The “unreasonable effectiveness” of deep learning is being tempered by the need for physical consistency. Researchers are embedding physical laws and geometric constraints directly into neural architectures to bridge the “Representation-Action Gap.”

• Neural Operators and Physics: Fractional Laplace Neural Operators: Exact Architectures, an Expressivity Frontier at Criticality, and Certified Stability for Memory-Driven Network Dynamics and Learning ab initio phase-field models embed physical laws into operators. In robotics, PACT: End-to-End Learning of Human Pose, Contacts, and Forces from Video and TANGO: Humanoid Navigation in Cluttered Environments with a Whole-Body Vision-Language-Action Model ensure motion is consistent with interaction forces.
• World Modeling: Oneira: From Open-Ended Generation to Open-World Interaction in Video World Models and 4Director: Controlling Video World Models with Rigid 3D Geometry demonstrate that persistent state and rigid SE(3) trajectories are essential for grounding AI in the physical world.
• Diagnostic Benchmarking: VTR-Bench: A Systematic Benchmark for Evaluating Visual Text Rendering in Video Generation and DiVid: Diagnosing Dimension-Specific Diversity Collapse in Video Generation Models allow us to pinpoint exactly where models fail, moving beyond aggregate metrics.

Theme 4: Optimization, Efficiency, and Adaptation

As models grow, the cost of training and inference has become a primary constraint, leading to innovations in memory-efficient optimizers and parameter-efficient adaptation.

• Optimizer Innovations: AF-Muon: An AdamW-Free Muon Optimizer for Tied-Embedding Models and TACO: Ternary Absolute-max Column-wise One-sparse Optimizer for LLM Fine-Tuning rethink update rules to reduce memory footprints.
• Inference-Time Scaling: How Much Can Language Models Gain from Test-Time Computation? and CAST: Cost-Aware Speculative Trees from One-Pass Block Drafters explore the trade-offs of spending compute at inference time.
• Parameter-Efficient Adaptation: Platonic Task Arithmetic treats model updates as composable objects, while Harnessing Domain Specialists in Multimodal Mixture-of-Experts for Efficient Adaptation leverages semantic modularity for efficient fine-tuning.
• Concept Erasure: RASteer: Retain-Aware Activation Steering for Concept Erasure in Diffusion Models and BARRIER: Bounded Activation Regions for Robust Information Erasure provide principled, training-free methods to remove unwanted concepts while maintaining model integrity.

Theme 5: Domain-Specific Foundations and Scientific Discovery

AI is increasingly acting as a “co-scientist,” capable of handling longitudinal data and complex reasoning in high-stakes domains like biology, medicine, and law.

• Scientific Discovery: Higher-Order Molecular Grammars for Generative and Foundation Models in Chemistry and EurekaBench: Measuring Agentic Ability to Discover New Scientific Insights demonstrate how structural priors and agentic workflows lead to genuine scientific insights.
• Clinical and Legal Rigor: Scaling Clinical Judgment to Evaluate Medical AI and JusticeAxis: Benchmarking Legal Judgment between Rigid Rule Application and Ungrounded Discretion emphasize the need for reference-anchored tasks in regulated fields.
• Trustworthiness: The Hidden Costs of 99% Accuracy: A Trustworthiness Audit of the Telco Customer Churn Benchmark serves as a vital reminder that high accuracy on benchmarks can mask fundamental failures in pipeline design and calibration.